Serum raman spectrum non-invasive pathological detection device
By combining discrete wavelet transform and deep learning models, the problems of insufficient spectral quality control and model interpretability in Raman spectroscopy analysis are solved, enabling efficient and reliable early cancer diagnosis and improving the accuracy and stability of detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TSINGHUA UNIVERSITY
- Filing Date
- 2025-09-29
- Publication Date
- 2026-07-21
AI Technical Summary
In existing technologies, Raman spectroscopy analysis in early cancer diagnosis suffers from problems such as difficulty in controlling spectral quality, difficulty in identifying complex spectra, and insufficient model decision-making mechanisms and interpretability, resulting in insufficient detection sensitivity and reliability.
Discrete wavelet transform is used to perform multi-scale frequency decomposition of the spectrum. Deep learning model is used for feature extraction and interpretability analysis. Qualified spectral data are screened by a preset wavelet transform signal-to-noise ratio energy criterion. A pre-trained convolutional neural network is used for pathological discrimination. Interpretability analysis is combined to locate key intervals and characteristic peaks and output biomolecular attribution information.
It improves the accuracy and reliability of early cancer diagnosis, meets the needs of large-scale non-invasive clinical screening, and achieves efficient and reliable early cancer accurate diagnosis.
Smart Images
Figure CN121186010B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of Raman spectroscopy analysis and machine learning technology, and in particular to a non-invasive pathological detection device using serum Raman spectroscopy. Background Technology
[0002] Cancer is the second leading cause of death worldwide, after cardiovascular disease, and has become a major threat to public health. Studies show that the five-year survival rate of cancer patients is closely related to the pathological stage at diagnosis: taking gastric cancer as an example, if diagnosed at an early stage, the five-year survival rate can exceed 80%, while if diagnosed at an advanced stage, the survival rate drops sharply to below 20%. Currently, most cancer patients are diagnosed only after the disease has progressed to the middle or late stages, greatly limiting the timeliness and effectiveness of treatment interventions. Therefore, developing a detection method that can accurately diagnose precancerous lesions or early stages is of great significance for improving the early detection rate and treatment effectiveness of cancer, and reducing its mortality rate.
[0003] Early screening and diagnosis of cancer still face significant challenges under current conditions. Taking early gastrointestinal cancer screening as an example, the currently accepted best method is endoscopy, which allows direct observation of the tissue morphology of lesions to aid in diagnosis. However, this method heavily relies on the doctor's clinical experience and subjective judgment, and early cancerous changes are often accompanied by very subtle tissue morphological changes that are difficult to accurately identify with the naked eye, easily leading to missed or misdiagnosed cases. Furthermore, pathological biopsy, as the "gold standard" with the highest diagnostic accuracy, provides cellular-level pathological evidence, but it is an invasive procedure, causing significant patient discomfort, and the sampling, staining, and interpretation processes are time-consuming and labor-intensive, making it unsuitable as a large-scale initial screening method.
[0004] In contrast, liquid biopsies, due to their convenient sampling, lower cost, and greater safety, have the potential for application in high-frequency, large-scale population screening. Raman spectroscopy, as a "fingerprint" of substances, can obtain important information such as the composition and concentration distribution of substances in a sample at the molecular level, and has been widely used in recent years for high-precision characterization in liquid biopsies. However, its practical application still faces many challenges: biological samples are generally accompanied by strong fluorescence backgrounds, while the concentration of cancer-related substances in body fluids is extremely low, and their spontaneous Raman signals are weak and easily drowned out by fluorescence signals, resulting in low spectral signal-to-noise ratios and limited detection sensitivity. To address this, surface-enhanced Raman spectroscopy (SERS) was introduced, effectively overcoming the problem of strong fluorescence signals and weak spontaneous Raman signals in body fluids. However, while SERS enhances the target signal, it also simultaneously amplifies the Raman scattering of various substances in body fluids, resulting in a richer signal composition in the spectrum, more significant differences in peak broadening, and a substantial increase in spectral complexity, bringing new challenges to subsequent spectral analysis.
[0005] Traditional spectral analysis methods typically employ dimensionality reduction techniques (such as Principal Component Analysis (PCA)) combined with classification algorithms (such as Linear Discriminant Analysis (LDA) or Support Vector Machines (SVM)) to classify Raman spectral data. While this approach has some effectiveness, its performance is often limited when dealing with highly complex, highly nonlinear, and low signal-to-noise ratio spectral data. With the development of deep learning technology, researchers are attempting to transform spectral data into a format more suitable for image recognition models. For example, some methods acquire two-dimensional stimulated Raman scattering histology (SRH) images of lesion areas through scanning imaging, or construct two-dimensional feature maps of the spectrum using mathematical techniques such as spectral recursion and fast Fourier transform, and then input these maps into convolutional neural networks (CNNs) for classification and recognition. These methods fully utilize the mature structures of existing image recognition networks, achieving high classification accuracy. However, they also exacerbate the "black box" problem of machine learning models: if scanning imaging is used, it typically only presents signals from one or two peaks, resulting in the discarding of a large amount of original spectral information during the conversion process. Furthermore, if mathematical methods are used to generate two-dimensional feature maps, the physical connection between wavenumbers and biomolecular vibrational modes in the original spectrum is weakened. Ultimately, while the model possesses some recognition ability, it lacks physical interpretability, making it difficult to trace the specific Raman peak positions used in the model's judgment and the biological mechanisms they reflect, thus limiting its widespread clinical application.
[0006] Among related technologies, the existing SERS spectral data analysis process still has significant shortcomings in terms of spectral quality control, effective identification of complex spectra, and transparency and interpretability of model decision-making mechanisms. It is difficult to meet the dual requirements of accuracy and interpretability for early cancer spectral diagnosis, and these issues urgently need to be addressed. Summary of the Invention
[0007] This application provides a non-invasive pathological detection device for serum Raman spectroscopy to solve the problems of difficult control of spectral quality, difficulty in identifying complex spectra, and deficiencies in model decision-making mechanisms and interpretability in the prior art. It effectively mines potential cancer marker features in the spectrum and helps to achieve efficient and reliable early cancer accurate diagnosis.
[0008] To achieve the above objectives, the first aspect of this application proposes a non-invasive pathological detection device for serum Raman spectroscopy, comprising: an acquisition module, a screening module, a storage module, a calculation module, and an inference module.
[0009] The acquisition module is used to acquire the surface-enhanced Raman spectrum of the serum sample to be tested and obtain the spectral data to be analyzed.
[0010] The filtering module is used to calculate the quality score of the spectral data to be analyzed based on a preset wavelet transform signal-to-noise ratio energy criterion, and based on the quality score, filter out data in the spectral data to be analyzed that are less than a preset score to obtain qualified spectral data.
[0011] A storage module is used to store the qualified spectral data into a preset database based on a preset storage format;
[0012] The calculation module is used to read qualified spectral data from the preset database and calculate and extract feature signals containing significant discriminative features based on the qualified spectral data.
[0013] The inference module is used to input the feature signal into a pre-trained convolutional neural network model, output the inference process and pathological discrimination result of the spectral data to be analyzed, and based on interpretability analysis, locate the key interval and feature peak corresponding to the pathological discrimination result, and output the biomolecular attribution information of the corresponding Raman feature peak according to the key interval or the feature peak.
[0014] According to one embodiment of this application, the acquisition module includes:
[0015] The acquisition unit is used to acquire the surface-enhanced Raman spectrum of the serum sample to be tested;
[0016] The processing unit is used to process the surface-enhanced Raman spectrum based on a unified wavenumber range and an interpolation resampling strategy to obtain the spectral data to be analyzed.
[0017] According to one embodiment of this application, the computing module includes:
[0018] The reading unit is used to read the qualified spectral data and the pre-trained convolutional neural network model from the preset database;
[0019] The preprocessing unit is used to perform standardized preprocessing on the qualified spectral data to obtain preprocessed spectral data;
[0020] The decomposition unit is used to perform multi-scale frequency decomposition on the preprocessed spectral data using a preset discrete wavelet transform strategy to obtain candidate effective feature signals.
[0021] The feature extraction unit is used to extract feature signals containing significant discriminative features from the candidate valid feature signals based on a preset frequency feature combination strategy.
[0022] According to one embodiment of this application, the preprocessing unit is specifically used for:
[0023] The normalization subunit is used to normalize the qualified spectral data using a preset Z-score standardization to obtain normalized spectral data.
[0024] The noise reduction and suppression subunit is used to perform noise removal and background suppression processing on the normalized spectral data to obtain the preprocessed spectral data.
[0025] According to one embodiment of this application, the decomposition unit is specifically used for:
[0026] A preset discrete wavelet transform strategy is used to perform multi-scale frequency decomposition on the preprocessed spectral data to obtain the first decomposition result.
[0027] According to one embodiment of this application, the serum Raman spectroscopy non-invasive pathological detection device further includes: a training module, the training module comprising:
[0028] The sample acquisition unit reads spectral data samples from the preset database, preprocesses the spectral data samples, and obtains feature signal samples based on the preprocessed data samples.
[0029] A partitioning unit is used to divide the feature signal samples into a training set, a validation set, and a test set based on a preset partitioning strategy.
[0030] The training unit constructs a convolutional neural network and inputs the training set into the convolutional neural network to train and obtain initial model parameters;
[0031] The validation unit is used to input the validation set into the convolutional neural network for performance evaluation based on the initial model parameters, and adjust the initial model parameters according to the performance evaluation results until the loss function of the validation set converges to obtain the optimal model parameters.
[0032] The testing unit is used to input the test set into the convolutional neural network for model testing based on the optimal model parameter combination, and obtain the pre-trained convolutional neural network model when the test results meet the preset requirements.
[0033] According to one embodiment of this application, the dividing unit is specifically used for:
[0034] Based on a preset random sampling strategy, a portion of samples are extracted from the feature signals to construct the test set, and the remaining samples are divided into the training set and the validation set using K-fold cross-validation.
[0035] According to one embodiment of this application, the serum Raman spectroscopy non-invasive pathological detection device further includes:
[0036] The display module is used to visualize the reasoning process and the pathological discrimination results through a preset graphical user interface, enabling clinicians to make non-invasive pathological diagnoses based on the displayed results.
[0037] According to one embodiment of this application, the screening module includes:
[0038] The extraction unit is used to perform discrete wavelet decomposition on the spectral data to be analyzed, and extract decomposed signals of different frequency bands based on the second decomposition result.
[0039] The calculation unit is used to calculate the ratio of the effective feature energy of the first frequency band to the noise signal energy of the second frequency band based on the decomposed signals of the different frequency bands, so as to obtain the quality fraction of the spectral data to be analyzed, wherein the frequency level of the first frequency band is determined by a preliminary experiment, and the second frequency band is the highest frequency decomposed signal;
[0040] The filtering unit is used to filter out data in the spectral data to be analyzed that are less than a preset score threshold based on the quality score, so as to obtain the qualified spectral data.
[0041] According to one embodiment of this application, the interpretability analysis is based on gradient-weighted class activation mapping and spectral difference analysis, wherein the gradient-weighted class activation mapping is used to adaptively locate key spectral frequency bands in the discrimination process of the convolutional neural network, and the spectral difference analysis is used to find Raman characteristic peaks and their biomolecular attribution information corresponding to the key spectral frequency bands.
[0042] The non-invasive pathological detection device for serum Raman spectroscopy proposed in this application collects surface-enhanced Raman spectra of the serum to be tested, obtains the spectral data to be analyzed, filters qualified spectral data based on a preset wavelet transform signal-to-noise ratio energy criterion, and stores it in a database. It reads qualified data and extracts feature signals, inputs the signals into a pre-trained model, outputs the model's reasoning process and pathological diagnosis results, and, combined with interpretability analysis, locates key regions and characteristic peaks, outputting the corresponding biomolecular attribution information. This solves the problems of difficult-to-control spectral quality, difficult-to-analyze complex spectra, and deficiencies in model decision-making mechanisms and interpretability in existing technologies. It effectively mines potential cancer marker features in the spectrum, contributing to efficient and reliable early cancer diagnosis.
[0043] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description
[0044] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:
[0045] Figure 1 This is a block diagram of a non-invasive pathological detection device for serum Raman spectroscopy provided according to an embodiment of this application;
[0046] Figure 2 This is a schematic diagram of a spectral database structure provided according to an embodiment of this application;
[0047] Figure 3 This is a schematic diagram of the spectral frequency decomposition process based on wavelet transform according to Embodiment 1 of this application;
[0048] Figure 4 This is a histogram of the mass fraction distribution of a set of spontaneous Raman scattering spectral data provided according to Embodiment 1 of this application;
[0049] Figure 5 This is a schematic diagram illustrating the dataset partitioning process and K-fold cross-validation principle according to an embodiment of this application;
[0050] Figure 6 This is a block diagram of a non-invasive pathological detection device for serum Raman spectroscopy according to an embodiment of this application;
[0051] Figure 7 This is a graph showing the relationship between the recognition accuracy and feature extraction width of different frequency features according to Embodiment 2 of this application;
[0052] Figure 8 This is a comparative analysis diagram of the differences between the mid-frequency band features D6 and D7 provided in Embodiment 2 of this application and the original spectrum after piecewise linear baseline correction;
[0053] Figure 9 A visualization and interpretation diagram of the key intervals discriminated by the model obtained by combining gradient-weighted class activation mapping technology and spectral difference analysis method according to Embodiment 2 of this application;
[0054] Figure 10 This is a graphical user interface diagram illustrating the integrated multi-functional algorithm provided according to Embodiment 2 of this application. Detailed Implementation
[0055] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.
[0056] To address the shortcomings of existing non-invasive serum pathological detection technologies based on surface-enhanced Raman spectroscopy, such as insufficient spectral quality control, significant influence from fluorescence background and noise, difficulty in effectively identifying complex spectral features, and the uninterpretable nature of deep learning models ("black box"). This application aims to provide a non-invasive serum Raman spectroscopy pathological detection device. This device introduces Discrete Wavelet Transform (DWT) to perform multi-scale frequency decomposition of the spectrum, enabling pre-screening and standardized preprocessing of the spectrum. Combined with a deep learning model, it improves the accuracy of pathological discrimination. Furthermore, based on the model's interpretability, it visualizes the discrimination criteria and analyzes biomolecular attribution, thereby enhancing the accuracy, stability, and reliability of early cancer detection and meeting the needs of large-scale non-invasive clinical screening.
[0057] The serum Raman spectroscopy non-invasive pathological detection device according to the embodiments of this application will be described below with reference to the accompanying drawings.
[0058] Figure 1 This is a schematic diagram of a serum Raman spectroscopy non-invasive pathological detection device according to an embodiment of this application.
[0059] like Figure 1 As shown, the serum Raman spectroscopy non-invasive pathological detection device 10 includes: an acquisition module 100, a screening module 200, a storage module 300, a calculation module 400, and an inference module 500.
[0060] The acquisition module 100 is used to acquire the surface-enhanced Raman spectrum of the serum sample to be tested and obtain the spectral data to be analyzed.
[0061] Optionally, in some embodiments, the acquisition module 100 includes: an acquisition unit 101 for acquiring the surface-enhanced Raman spectrum of the serum sample to be tested; and a processing unit 102 for processing the surface-enhanced Raman spectrum based on a unified wavenumber range and an interpolation resampling strategy to obtain interpolated spectral data to be analyzed.
[0062] Serum samples to be tested refer to representative biological samples isolated and prepared from the blood of organisms, whose chemical composition, biomarkers, or abnormal substances need to be analyzed using specific detection technologies. Surface-enhanced Raman spectroscopy is an analytical technique based on the Raman scattering effect, which significantly enhances the Raman signal of molecules through special metal nanostructures. A unified wavenumber interval refers to a fixed and consistent Raman shift (wavenumber) range standard pre-defined to eliminate differences in the spectral wavenumber dimension caused by different detection conditions. Interpolation resampling strategy refers to a systematic method that uses a unified wavenumber interval as the target wavenumber axis and a specific mathematical interpolation algorithm to transform the differentially distributed wavenumber-signal data points in the original SERS spectrum into a set of regularly consistent and equally spaced data points within the unified wavenumber interval.
[0063] Specifically, the acquisition unit 101 acquires the surface-enhanced Raman spectrum of the serum sample to be tested; then the processing unit 102 performs preliminary processing on the original spectrum: based on the Raman characteristics of the biomolecule structure in the serum, a suitable uniform wavenumber range is preset, invalid data outside the range is eliminated, and combined with the preset equally spaced target wavenumber grid, an interpolation resampling is performed using an interpolation algorithm adapted to the spectral signal restoration requirements, and finally a measurement spectrum with consistent wavenumber dimension and uniform data format is obtained, so as to eliminate systematic errors caused by instrument calibration deviation, batch detection parameter fluctuations, etc., and obtain a measurement spectrum with uniform format.
[0064] The screening module 200 is used to calculate the quality score of the spectral data to be analyzed based on the preset wavelet transform signal-to-noise ratio energy criterion, and based on the quality score, screen out data in the spectral data to be analyzed that are less than the preset score threshold to obtain qualified spectral data.
[0065] The preset wavelet transform signal-to-noise ratio energy criterion is a pre-defined spectral data quality assessment standard based on wavelet transform technology. The quality score is a numerical indicator used to quantitatively evaluate the quality of a single spectral data sample. The preset score threshold is a pre-set threshold used to determine whether the quality score of a single spectrum meets the standard before quality screening of the spectral data.
[0066] Optionally, in some embodiments, the screening module 200 includes: an extraction unit 201, used to perform discrete wavelet decomposition on the spectral data to be analyzed, and extract decomposed signals of different frequency bands according to the second decomposition result; a calculation unit 202, used to calculate the ratio of the effective feature energy of the first frequency band to the noise signal energy of the second frequency band based on the decomposed signals of different frequency bands, to obtain the quality score of the spectral data to be analyzed, wherein the frequency level of the first frequency band is determined by a pre-experiment, and the second frequency band is the highest frequency decomposed signal; and a screening unit 203, used to screen out data in the spectral data to be analyzed that are less than a preset score threshold based on the quality score, to obtain qualified spectral data.
[0067] Discrete wavelet decomposition (DWD) refers to decomposing the original signal into multiple sub-signals of different frequency levels using wavelet basis functions. This separates the slowly varying trends and rapidly varying details in the signal, ultimately yielding structured signal components that are easier to analyze and process. The second decomposition result refers to the specific set of signals obtained after the second decomposition step in the multi-scale iterative process of discrete wavelet decomposition. Decomposing the signal means breaking down the original signal, which contains a mixture of various information, into multiple sub-signals with distinct characteristics. Effective feature energy refers to the energy value corresponding to the portion of the signal carrying useful information extracted from the decomposed signal in signal processing. Noise signal energy refers to the energy value corresponding to the portion of the signal carrying useless interference information (i.e., noise) extracted from the decomposed signal in signal processing.
[0068] Specifically, the extraction unit 201 performs discrete wavelet decomposition on the spectral data to be analyzed. Through multi-scale iterative decomposition, it obtains a second decomposition result containing low-frequency approximate signals and high-frequency detail signals, and extracts decomposition signals of different frequency bands from this result, achieving separation of different frequency components in the spectrum. Based on this, the calculation unit 202 performs energy analysis on the extracted multi-band decomposition signals. The first frequency band decomposition signal, determined in a pre-experimental study, is defined as the main carrier of effective features, and its corresponding effective feature energy is calculated. The second frequency band decomposition signal, with the highest frequency, is defined as the main carrier of noise, and its corresponding noise energy is calculated. The ratio of these two values yields the quality score of the spectral data to be analyzed. The screening unit 203 uses a preset score as a quality screening threshold, compares the quality scores of all spectral data to be analyzed, and removes low-quality data with scores below the threshold (i.e., data whose high noise content may interfere with subsequent analysis). It retains qualified spectral data with acceptable scores and clear effective features, ensuring that the spectral data used in subsequent analyses has high reliability and effectiveness.
[0069] Storage module 300 is used to store qualified spectral data into a preset database based on a preset storage format.
[0070] The preset storage format refers to a set of rules that are set in advance to standardize the organization and storage structure of qualified spectral data in order to achieve standardized management, efficient reading and writing, and convenient access to subsequent modules.
[0071] Specifically, qualified spectral data are stored in a spectral database (i.e., a pre-defined database) according to a preset format. For example, such as... Figure 2 As shown, Figure 2This is a schematic diagram of a spectral database structure provided according to an embodiment of this application. The diagram contains a three-level hierarchy: "sample-measurement point-measurement spectrum": one sample corresponds to multiple measurement points, and one measurement point corresponds to multiple measurement spectra. Each sample includes attributes such as sample number, patient name, age, sampling site, sampling time, submission time, and pathological diagnosis result; each measurement point includes a measurement point number and pathological slide result (serving as the "gold standard" label for model training); each measurement spectrum includes attributes such as spectral number, Raman shift, signal intensity, measurement interval, laser wavelength, and integration time.
[0072] The calculation module 400 is used to read qualified spectral data from a preset database and calculate and extract feature signals containing significant discriminative features based on the qualified spectral data.
[0073] Optionally, in some embodiments, the screening module 400 includes: a reading unit 401, used to read qualified spectral data and a pre-trained convolutional neural network model from a preset database; a preprocessing unit 402, used to preprocess the qualified spectral data to obtain preprocessed spectral data; a decomposition unit 403, used to perform multi-scale frequency decomposition on the preprocessed spectral data using a preset discrete wavelet transform strategy to obtain candidate effective feature signals; and a feature extraction unit 404, used to screen feature signals containing significant discriminative features from the candidate effective feature signals based on a preset frequency feature combination strategy.
[0074] Among them, a feature signal with significant discriminative characteristics refers to a concise signal segment / information set extracted from qualified spectral data that possesses the core ability to stably distinguish different sample categories. A candidate effective feature signal refers to a set of multi-frequency signals that may contain effective feature information. A pre-defined frequency feature combination strategy refers to a systematic method that is pre-set to analyze the frequency attributes and discriminative capabilities of candidate effective feature signals, thereby accurately extracting core signals containing significant discriminative characteristics.
[0075] Specifically, the reading unit 401 reads spectral data (i.e., qualified spectral data) for model training and analysis from a preset database, and performs quality control on the qualified spectral data as needed, such as increasing the pre-screened quality score threshold. The preprocessing unit 402 preprocesses the qualified spectral data, such as normalization, noise reduction, and fluorescence suppression. The decomposition unit 403 performs multi-scale decomposition on the preprocessed spectral signal using discrete wavelet transform to obtain detail signals and approximate signals (i.e., candidate effective feature signals) at different frequency levels. The feature extraction unit 404 extracts feature signals containing significant discriminative features from the candidate effective feature signals based on a preset frequency feature combination strategy.
[0076] Optionally, in some embodiments, the preprocessing unit 402 is specifically used for: a normalization subunit, used to normalize the qualified spectral data using a preset Z-score normalization to obtain normalized spectral data; and a noise reduction and suppression subunit, used to perform noise removal and background suppression processing on the normalized spectral data to obtain preprocessed spectral data.
[0077] Background suppression refers to the preprocessing operation that uses specific algorithms or techniques to identify and reduce / eliminate background signal interference generated during spectral measurement, thereby highlighting the target analysis object.
[0078] Specifically, the preprocessing unit 402 preprocesses the qualified spectral data through two sub-units: the normalization sub-unit uses a preset Z-score standardization method to eliminate amplitude differences caused by instrument sensitivity fluctuations and sample concentration deviations, while fully preserving the relative intensity and peak position relationship of characteristic peaks to obtain normalized spectral data; the noise reduction and suppression sub-unit performs noise removal and background suppression processing on the normalized spectral data, using methods such as removing high-frequency noise and low-frequency background fluorescence to achieve noise reduction and fluorescence suppression, and finally outputs spectral data with clear features and low interference.
[0079] Optionally, in some embodiments, the decomposition unit 403 is specifically used to: perform multi-scale decomposition on the preprocessed spectral data to obtain a first decomposition result.
[0080] Specifically, the decomposition unit 403 uses a preset discrete wavelet transform strategy to perform multi-scale decomposition on the preprocessed spectral data, breaking down the spectral signal into a set of sub-signals covering the entire frequency band from low to high frequency, i.e., the first decomposition result. This result includes the fluorescence signal component in the lowest frequency band (smoothly varying background interference), the potential effective feature signal in the middle frequency band, and the high-frequency noise component in the highest frequency band (random interference with irregular and drastic fluctuations). Subsequently, a targeted mid-frequency band feature combination strategy is used to extract feature signals with significant discriminative characteristics. It should be noted that the first decomposition result and the second decomposition result are the same set of sub-signals, and their decomposition results are essentially the same. There is no order or difference in importance between them; they are only used to extract mid-frequency band feature combinations for input into the model and to calculate the signal-to-noise ratio quality score to screen out low-quality spectra, respectively.
[0081] The inference module 500 is used to input feature signals into a pre-trained convolutional neural network model, output the inference process and pathological discrimination results of the spectral data to be analyzed, and based on interpretability analysis, locate the key intervals and feature peaks corresponding to the pathological discrimination results, and output the biomolecular attribution information of the corresponding Raman feature peaks according to the key intervals and feature peaks.
[0082] Optionally, in some embodiments, interpretability analysis is based on gradient-weighted class activation mapping and spectral difference analysis, wherein gradient-weighted class activation mapping is used to adaptively locate the spectral bands that play a key role in the discrimination process of the convolutional neural network, and spectral difference analysis is used to find Raman characteristic peaks corresponding to the key spectral bands and their biomolecular attribution information.
[0083] Among these, pathological discrimination results refer to the definitive conclusions about the pathological attributes of the sample corresponding to the analyzed spectral data, obtained by learning the mapping relationship between feature signals and the pathological state of the sample. Interpretability analysis refers to the analytical method that breaks down the black box barrier in the decision-making process of a deep learning model through specific techniques, revealing how the model derives the final pathological discrimination result based on input features. The key interval refers to a specific continuous wavenumber range within the selected analytical band of Raman spectroscopy that significantly contributes to the discrimination of pathological states and is heavily relied upon by the analyzed model. Feature peaks refer to discrete spectral signal peaks within the key interval that have clear molecular vibrational significance and reflect changes in biomolecules related to the pathological state. Biomolecular attribution information refers to information that establishes a clear correspondence between identified feature peaks in Raman spectroscopy and specific biomolecules. Gradient-weighted class activation mapping is an interpretability technique for deep learning models.
[0084] Specifically, the inference module 500 receives the feature signal output from the feature extraction unit 404 and inputs the feature signal into the pre-trained convolutional neural network model. After the convolutional neural network model is trained on a spectral sample dataset labeled with pathological state (such as a normal / cancer serum Raman spectrum dataset), it has established a mapping relationship between the feature signal and the pathological state. Based on the input feature signal, it can simultaneously output two core pieces of information: first, the complete inference process from feature input to pathological conclusion generation of the spectral data to be analyzed (including the feature extraction and integration logic of each layer of the model); and second, the pathological discrimination result of the sample corresponding to the spectral data to be analyzed (including the specific pathological state category and discrimination confidence, such as: "Pathological state: early gastric cancer; confidence: 94.5%").
[0085] Furthermore, to ensure the credibility and traceability of the pathological diagnosis results, the inference module 500 is also configured to perform reverse analysis of the discrimination process of the convolutional neural network model based on interpretability analysis technology: by interpretability analysis, the key spectral ranges and characteristic peaks directly related to the current pathological diagnosis results are located, and based on the established Raman spectroscopy-biomolecule association database, the molecular vibration modes of the key ranges and characteristic peaks are matched, thereby outputting the biomolecule attribution information corresponding to the molecular structure.
[0086] Optionally, interpretability analysis is based on gradient-weighted class activation mapping (Grad-CAM) and spectral difference analysis. Grad-weighted class activation mapping calculates the importance score of the signal at each wavenumber point in the model identification process, and then adaptively highlights the spectral regions that the model focuses on during the discrimination process. Spectral difference analysis uses the shaded curve of the mean and standard deviation of key spectral bands as the analysis object. By comparing the signal differences between the spectrum of the cancerous area and the normal spectrum in this band, it screens out spectral intervals with significant differences before and after cancer. It then combines Raman spectroscopy databases to complete the biomolecular attribution matching of characteristic peaks, and finally outputs complete interpretive information including key spectral intervals, corresponding Raman characteristic peaks, and biomolecular attribution.
[0087] To facilitate further understanding of the serum Raman spectroscopy non-invasive pathological detection device of this application by those skilled in the art, an embodiment 1 of this application provides a spectral quality pre-screening method based on wavelet transform signal-to-noise ratio energy criterion, aiming to verify the applicability and effectiveness of the spectral quality control method in this application. The specific method is as follows:
[0088] In Example 1, the experimental data used were spontaneous Raman scattering spectra of early gastric cancer tissue collected independently, totaling 355 spectra, including 147 normal spectra and 208 cancerous spectra. This dataset differs from the surface-enhanced Raman spectroscopy data of the serum samples to be tested, and its main purpose is to demonstrate that the quality control method proposed in this application is also applicable to spontaneous Raman spectroscopy.
[0089] This embodiment performs Z-score normalization on the original spectrum to eliminate amplitude differences between spectra. Subsequently, discrete wavelet transform is used to perform an 11-level decomposition of the normalized Raman spectrum, employing the Daubechies 8 (db8) wavelet basis functions. Figure 3 As shown, Figure 3 This diagram illustrates a wavelet transform-based spectral frequency decomposition process provided in Embodiment 1 of this application. The left side of the diagram shows the original spectrum, and the right side shows the obtained wavelet sub-signals at each stage, where D1 to D... 11 For detailed signals at different frequency levels, A 11 This is the lowest frequency approximation signal. As can be seen from the algorithm principle of wavelet transform, the original signal can be completely reconstructed by superimposing all sub-signals.
[0090] To evaluate the impact of features contained in sub-signals of different frequencies on classification performance, a convolutional neural network model with a fixed structure was used to train each sub-signal layer separately. This model consists of one convolutional layer and three fully connected layers. The training hyperparameters were set as follows: 20 training epochs, batch size 8, initial learning rate 0.001, and L2 regularization coefficient 1×10⁻⁶. -4 .
[0091] Based on the preliminary experimental results above, high-frequency signals (such as D1) mainly contain noise components, while mid-frequency signals (such as D6) are better able to reflect Raman characteristic information with discriminative effect. Therefore, it is proposed that the energy ratio of the mid-frequency signal D6 to the high-frequency signal D1 be used as a fraction of spectral quality to measure the intrinsic quality of the measured spectrum, defined as follows:
[0092] ;
[0093] in, E 6 represents the energy of the intermediate frequency signal D6. E 1 represents the energy of the high-frequency signal D1. x 6 represents the discrete sampling sequence of the intermediate frequency signal D6. x 1 represents the discrete sampling sequence of the high-frequency signal D1.
[0094] like Figure 4 As shown, Figure 4 This is a histogram of the mass fraction distribution of a set of spontaneous Raman scattering spectral data provided according to Embodiment 1 of this application. This figure shows the mass fraction distribution corresponding to the spontaneous Raman scattering spectral data of the tissue obtained above. It can be observed that, due to the relatively weak spontaneous Raman scattering signal of biological samples, the overall spectral quality of this dataset is low, with a large amount of spectral data exhibiting low signal-to-noise ratios.
[0095] Without any quality control, directly extracting features from all spectra and inputting them into a convolutional neural network model for classification results in an accuracy of only 81.9%. However, when the pre-screening threshold for spectral quality score is set to 0.5, the number of effective spectral samples after pre-screening decreases from the original 355 to 214, while keeping other training parameters unchanged. This significantly improves the model's recognition accuracy to 89.3%. Therefore, removing some low-quality data and retaining high-quality spectra significantly enhances the model's discriminative ability and robustness.
[0096] Example 1 of this application addresses a set of Raman spectral data with generally low quality. By setting a reasonable quality pre-screening threshold (e.g., 0.5), a nearly 10% improvement in recognition accuracy can be achieved while sacrificing approximately one-third of the sample quantity. This effectively demonstrates the practical application value of the quality control method based on wavelet transform signal-to-noise ratio energy criterion proposed in this application. Furthermore, while increasing the quality pre-screening threshold is expected to yield a higher-quality training set, thereby improving model performance, a trade-off must be struck between sample quantity and quality. Blindly increasing the pre-screening threshold will lead to a significant decrease in the number of effective training samples, potentially affecting the model's generalization ability. Therefore, in practical applications, the optimal spectral pre-screening strategy should be selected based on the characteristics of the dataset, comprehensively considering both quality and quantity factors.
[0097] It should be noted that this quality pre-screening criterion is based on the essential frequency structure characteristics of Raman spectral signals, and has good universality and scalability. It is not only applicable to the field of early cancer detection in this application, but also provides a generalizable quality control method for other types of spectral analysis tasks.
[0098] Therefore, surface-enhanced Raman spectra of the serum samples are collected to obtain spectral data for analysis. Qualified spectral data are obtained based on a pre-defined wavelet transform signal-to-noise ratio energy criterion and stored in a database. Qualified data are read, and feature signals are extracted. These signals are input into a pre-trained model, which outputs the reasoning process and pathological results. Interpretability analysis is used to locate key regions and characteristic peaks, and the corresponding biomolecular attribution information is output. This approach solves the problems of difficult-to-control spectral quality, difficulty in analyzing complex spectra, and deficiencies in model decision-making mechanisms and interpretability in existing technologies. It effectively mines potential cancer marker features in the spectrum, contributing to efficient and reliable early cancer diagnosis.
[0099] Furthermore, the pre-trained convolutional neural network model is the core foundation of pathological discrimination reasoning, directly determining the ability to analyze key regions and biomolecule attributions during the reasoning stage. The training implementation scheme of this model will be further elaborated below.
[0100] Optionally, in some embodiments, the serum Raman spectroscopy non-invasive pathological detection device 10 further includes: a training module 600, which includes: a sample acquisition unit 601, which reads spectral data samples from a preset database, preprocesses the spectral data samples, and obtains feature signal samples based on the preprocessed data samples; a partitioning unit 602, which partitions the feature signal samples into a training set, a validation set, and a test set based on a preset partitioning strategy; a training unit 603, which constructs a convolutional neural network, inputs the training set into the convolutional neural network for training until the model loss function converges, and obtains initial model parameters; a validation unit 604, which inputs the validation set into the convolutional neural network for performance evaluation based on the initial model hyperparameters, and adjusts the model hyperparameter combination according to the performance evaluation results to maximize the accuracy of the model on the validation set and obtain the optimal hyperparameter combination; and a testing unit 605, which inputs the test set into the convolutional neural network for model testing based on the optimal model parameter combination, and obtains the pre-trained convolutional neural network model when the test results meet preset requirements.
[0101] Optionally, in some embodiments, the partitioning unit 602 is specifically used to: extract a portion of samples from the feature signals to construct a test set based on a preset random sampling strategy, and use K-fold cross-validation to partition the remaining samples into a training set and a validation set.
[0102] The training set refers to the core dataset used for learning the parameters of the convolutional neural network model, partitioned from the feature signal samples. The validation set refers to the intermediate evaluation dataset separated from the feature signal samples. The test set refers to the independent final evaluation dataset separated from the feature signal samples. Initial model parameters refer to the set of network parameters initially generated during the first training phase of the convolutional neural network by learning the mapping relationship between the feature signal samples in the training set and the pathological labels. Model loss function convergence refers to the state where, through iterative adjustment of the initial model parameters, the loss function value of the model on the training and validation sets gradually decreases and tends to stabilize, and the change amplitude is less than a preset threshold in consecutive iterations. The optimal model hyperparameter combination refers to the final combination of convolutional neural network hyperparameters determined by maximizing the accuracy of the validation set, provided that the loss function on the validation set converges.
[0103] Specifically, in the training module 600, the sample acquisition unit 601 reads spectral data samples from a preset database, preprocesses the spectral data samples, and obtains feature signal samples based on the preprocessed data samples. It should be noted that the implementation method of this sample acquisition unit 601 is the same as that of the calculation module 400, and will not be described in detail here.
[0104] The partitioning unit 602 partitions the feature signal dataset, dividing the feature signal samples into a training set, a validation set, and a test set based on a preset partitioning strategy. For example, ... Figure 5 As shown, Figure 5 This diagram illustrates the dataset partitioning process and K-fold cross-validation principle according to an embodiment of this application. In this diagram, a subset of samples is first randomly selected to construct an independent external test set. The remaining samples are then divided into a training set and a validation set using K-fold cross-validation, and input into the convolutional neural network model. K-fold cross-validation refers to randomly dividing the samples into K subsets of similar size. In each iteration, one subset is selected as the validation set, and the remaining subsets are used for model training. This process is repeated until each subset has been used as a validation set in a model evaluation, thereby improving data utilization and model reliability. In training unit 603, the training and validation sets are input into the convolutional neural network model. In validation unit 604, the structure and hyperparameters of the convolutional neural network model are optimized and adjusted based on the average accuracy of the validation set. In testing unit 605, after optimization, the model's performance and generalization ability are finally evaluated using an independent test set to ensure that the model performs well on unseen data, and the pre-trained convolutional neural network model is obtained and saved.
[0105] This provides core parameter support for the accurate pathological identification of the spectrum to be detected and for subsequent interpretability analysis.
[0106] Furthermore, in order to transform the model's reasoning logic and judgment results into clinically interpretable information, and to help doctors quickly understand the basis of the test and make accurate judgments, the device is also equipped with a display module.
[0107] Optionally, in some embodiments, the blood Raman spectroscopy non-invasive pathological detection device 10 further includes a display module 700, which is used to visualize the reasoning process and pathological discrimination results through a preset graphical user interface, so that clinicians can make non-invasive pathological diagnoses based on the displayed results.
[0108] The pre-designed graphical user interface (GUI) refers to a visually interactive interface that is pre-designed and fixed based on the operating habits of clinicians and the professional information needs of spectral pathology testing. Visualization refers to the technical process of transforming the abstract reasoning process of convolutional neural network models and specific pathological judgment results into intuitive and quickly interpretable graphics, images, or dynamic interactive forms through the pre-designed GUI. Non-invasive pathological diagnosis refers to a pathological diagnostic method that determines the presence and type of lesions in the body through model analysis and clinical interpretation, without damaging the integrity of the body's tissues or causing invasive procedures.
[0109] Specifically, the entire spectral analysis process constructed in this application embodiment can be integrated into a graphical user interface, improving the usability and deployment efficiency of the method. This interface includes functional modules such as spectral data input, quality control, preprocessing and feature extraction, model training, and exogenous spectral discrimination. Based on the reasoning and clear conclusions displayed on the interface, combined with the patient's clinical background, doctors can complete a non-invasive pathological diagnosis.
[0110] In summary, as Figure 6 As shown, Figure 6 According to one embodiment of this application, a serum Raman spectroscopy non-invasive pathological detection device 10 is provided. The device 10 includes: an acquisition module 100, a calculation module 200, a storage module 300, a screening module 400, an inference module 500, a training module 600, and a display module 700.
[0111] Therefore, it can assist doctors in quickly completing non-invasive pathological diagnoses by combining clinical background, effectively improving diagnostic efficiency and the reliability of results.
[0112] To facilitate those skilled in the art to further understand the serum Raman spectroscopy non-invasive pathological detection device of the present application embodiments, the present application provides an embodiment 2, proposing a spectral intelligent "white-box" recognition model based on esophageal cancer serum.
[0113] Example 2 of this application uses serum samples from esophageal cancer patients as the research object and proposes a spectral intelligent "white-box" recognition model based on wavelet frequency division feature extraction to achieve high-precision classification and recognition of serum spectra. In Example 2, the experimental data used are 863 self-collected serum surface-enhanced Raman scattering spectra, including 376 normal spectra and 487 esophageal cancer spectra.
[0114] In Embodiment 2 of this application, the raw SERS spectra of serum are read from the constructed database, and the spectral data are preprocessed by standardization. Then, the standardized spectra are decomposed by wavelet transform to extract sub-signals of different frequency levels as features for model training. The standardization, wavelet basis function, and decomposition level processing methods in this embodiment are the same as those in Embodiment 1, and will not be repeated here.
[0115] Subsequently, K-fold cross-validation was used to divide the dataset corresponding to each frequency feature, and the data were then input into the convolutional neural network model for training. During the dataset partitioning process, 20% of the samples were randomly selected to construct an independent test set, and the remaining data were divided into training and validation sets using ten-fold cross-validation. The convolutional neural network model used in this embodiment contains one convolutional layer and three fully connected layers. The model training hyperparameters were optimized using the average accuracy of the validation set, and the final settings were as follows: 10 training cycles, batch size of 4, initial learning rate of 0.0001, and L2 regularization coefficient of 8 × 10⁻⁶. -4 .
[0116] After training, the classification performance of each model is comprehensively evaluated using metrics such as average accuracy, specificity, sensitivity, and receiver operating characteristic curves on independent test sets.
[0117] like Figure 7 As shown, Figure 7 This is a graph showing the relationship between the recognition accuracy and feature extraction width of different frequency features according to Embodiment 2 of this application. The horizontal axis of the graph represents the "signal extraction width" of each frequency feature, reflecting the degree of broadening of the feature at that frequency in the Raman spectrum. From left to right, the frequency levels gradually transition from high to low frequencies. It can be observed that the model recognition accuracy reaches 98.8% at position D7 in the mid-frequency band; and when training using the mid-frequency feature combination D6+D7, the recognition accuracy can be further improved to 99.0%. This indicates that the wavelet frequency decomposition method successfully extracted the rich and crucial Raman features of the original spectrum located in the mid-frequency band, significantly enhancing the classification performance of the model.
[0118] As a comparative implementation, Embodiment 2 of this application also introduces two traditional spectral analysis methods: one is classification based on principal component analysis (PCA) followed by dimensionality reduction and then linear discriminant analysis (LDA), and the other is classification based on principal component analysis (PCA) followed by dimensionality reduction and then support vector machine (SVM). Experimental results show that the recognition accuracies of these two methods on the dataset used in this example are 89.5% and 95.9%, respectively. In contrast, using the wavelet mid-frequency features extracted by the method of this application as model input and combining them with a convolutional neural network for classification significantly improves the recognition accuracy on the same dataset, verifying the effectiveness of the mid-frequency features in spectral classification tasks.
[0119] To explain the classification criteria of the model at the biomolecular level, Example 2 performed spectral difference analysis on wavelet features at different frequency levels. In high-frequency signals, the signal differences between normal and cancerous spectra were not significant. However, in the mid-frequency range (especially the D6 and D7 bands), the differences in spectral features between normal and cancerous samples increased significantly. Figure 8 As shown, Figure 8 This is a comparative analysis diagram of the differences between the mid-frequency band features D6 and D7 and the original spectrum after piecewise linear baseline correction, provided in Example 2 of this application. To clarify the positions of the difference peaks and their corresponding molecular information, Example 2 compares the spectral differences between the D6 and D7 bands (…). Figure 8 of (a), Figure 8 (b) shows the difference between the original spectrum and the spectrum after piecewise linear baseline correction. Figure 8 A comparative analysis was conducted on (c) to identify 11 Raman characteristic peaks that made significant contributions to the classification process. Figure 8 It is marked in (c).
[0120] To further explain the discrimination mechanism of the constructed model, Example 2 employs a gradient-weighted class activation mapping (GCAM) method to visualize and analyze the learning process within the convolutional neural network. Grad-CAM adaptively highlights the spectral regions that the model focuses on during the discrimination process. Figure 9 As shown, Figure 9 This is a visualization and interpretation diagram of the key intervals of the model obtained by combining gradient-weighted class activation mapping technology and spectral difference analysis method according to an embodiment of this application 2. Figure 9 (a) and Figure 9 (b) shows the average signal curves of normal and cancerous samples under D6 and D7 features, and the highlighted areas of their corresponding classifications.
[0121] according to Figure 9 (c) and Figure 9 As shown in (d), the region highlighted by Grad-CAM closely matches the key Raman peaks obtained from the aforementioned difference analysis based on wavelet frequency division features, indicating that the classification basis of convolutional neural networks has good interpretability at the molecular biological level. Furthermore, these two interpretative approaches are complementary: the Grad-CAM method helps discover more potential spectral sites that significantly contribute to classification, while the difference analysis of frequency features can verify the rationality and biological relevance of these new sites at the biomolecular level.
[0122] To improve the clinical usability of this method and further lower the deployment threshold of the algorithm, this embodiment 2 also developed a graphical user interface to integrate relevant analysis workflows, such as... Figure 10 As shown, Figure 10 This is a graphical user interface diagram illustrating a multi-functional algorithm integrated according to Embodiment 2 of this application. The graphical interface includes five active functional modules: spectral data reading, spectral preprocessing and wavelet transform frequency decomposition, spectral quality control, K-fold cross-validation model training, and exogenous spectrum prediction. It also includes auxiliary functional areas such as a top image display panel, a bottom status output bar, and a log information window. Figure 10 (a) shows the average Raman spectrum (mean ± standard deviation) after reading and preprocessing in Example 2. Figure 10 (b) is the mass fraction distribution histogram of the serum SERS spectral data; Figure 10 (c) shows the average receiver operating characteristic curve of the trained model on the independent test set during the K-fold cross-validation training process, which is used to evaluate the discriminative performance of the model; Figure 10 (d) represents the process of reasoning and discrimination performed by the model after inputting a single external spectral data.
[0123] It should be noted that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.
[0124] This enables effective quality control of SERS spectra in body fluids, improving data reliability. Starting from the essential structure of spectral data, this application utilizes wavelet transform to perform multi-scale decomposition of the original spectrum. Based on the energy characteristics of each frequency component, a quality score index is constructed to perform quality discrimination and pre-screening of the original spectrum, effectively eliminating low-quality data with low signal-to-noise ratios and significantly improving the accuracy and stability of subsequent analysis steps. Frequency domain separation enhances the classification performance of extracted features, significantly improving the recognition accuracy of the convolutional neural network model. This application leverages the natural frequency differences between Raman, fluorescence, and noise signals, employing wavelet transform to perform multi-scale frequency decomposition of the spectrum. Combined with a convolutional neural network model, the classification ability of signals in different frequency bands is evaluated, selecting the most discriminative mid-frequency band features for model construction, thereby effectively improving the specificity and sensitivity of early cancer identification. An interpretability analysis method is introduced, making the deep learning model "white-box." This application combines gradient-weighted class activation mapping technology with traditional spectral difference analysis methods to achieve a visual representation of the model's discrimination criteria. The examples demonstrate that the two methods have good consistency, can adaptively locate the key spectral regions that the model focuses on during the classification process, and find the corresponding molecular vibrational structures based on the position of the Raman peaks to infer potential biomolecular markers, providing an intuitive and reliable decision-making basis for medical and clinical diagnosis.
[0125] The non-invasive pathological detection device for serum Raman spectroscopy proposed in this application collects surface-enhanced Raman spectra of the serum to be tested, obtains the spectral data to be analyzed, obtains qualified spectral data based on a preset wavelet transform signal-to-noise ratio energy criterion, and stores it in a database. The qualified data is read and feature signals are extracted, and the signals are input into a pre-trained model, which outputs the reasoning process and pathological results. Combined with interpretability analysis, key regions and characteristic peaks are located, and the corresponding biomolecular attribution information is output. This solves the problems of difficult control of spectral quality, difficulty in analyzing complex spectra, and deficiencies in model decision-making mechanisms and interpretability in existing technologies. It effectively mines potential cancer marker features in the spectrum, contributing to efficient and reliable early cancer diagnosis.
[0126] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.
[0127] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0128] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. A non-invasive pathological detection device using serum Raman spectroscopy, characterized in that, include: The acquisition module is used to acquire the surface-enhanced Raman spectrum of the serum sample to be tested and obtain the spectral data to be analyzed. The filtering module is used to calculate the quality score of the spectral data to be analyzed based on a preset wavelet transform signal-to-noise ratio energy criterion, and based on the quality score, remove data in the spectral data to be analyzed that are less than a preset threshold to obtain qualified spectral data. A storage module is used to store the qualified spectral data into a preset database based on a preset storage format; The calculation module is used to read qualified spectral data from the preset database and extract feature signals containing significant discriminative features from the qualified spectral data based on wavelet transform. The inference module is used to input the feature signal into a pre-trained convolutional neural network model, output the inference process and pathological discrimination result of the spectral data to be analyzed, and based on interpretability analysis, locate the key interval and feature peak corresponding to the pathological discrimination result, and output the biomolecular attribution information of the corresponding Raman feature peak according to the key interval and the feature peak. The filtering module includes: an extraction unit for performing discrete wavelet decomposition on the spectral data to be analyzed and extracting decomposed signals of different frequency bands based on the second decomposition result; a calculation unit for calculating the ratio of effective feature energy of the first frequency band to noise signal energy of the second frequency band based on the decomposed signals of the different frequency bands, to obtain the quality score of the spectral data to be analyzed, wherein the frequency level of the first frequency band is determined by a pre-experiment, and the second frequency band is the highest frequency decomposed signal; and a filtering unit for filtering out data in the spectral data to be analyzed that are less than a preset score threshold based on the quality score, to obtain the qualified spectral data. The interpretability analysis is based on gradient-weighted class activation mapping and spectral difference analysis. The gradient-weighted class activation mapping is used to adaptively locate the key spectral frequency bands of the convolutional neural network in the discrimination process, and the spectral difference analysis is used to find the Raman characteristic peaks corresponding to the key spectral frequency bands and their biomolecular attribution information.
2. The apparatus according to claim 1, characterized in that, The acquisition module includes: The acquisition unit is used to acquire the surface-enhanced Raman spectrum of the serum sample to be tested; The processing unit is used to process the surface-enhanced Raman spectrum based on a unified wavenumber range and an interpolation resampling strategy to obtain the spectral data to be analyzed.
3. The serum Raman spectroscopy non-invasive pathological detection device according to claim 1, characterized in that, The computing module includes: The reading unit is used to read the qualified spectral data and the pre-trained convolutional neural network model from the preset database; The preprocessing unit is used to perform standardized preprocessing on the qualified spectral data to obtain preprocessed spectral data; The decomposition unit is used to perform multi-scale frequency decomposition on the preprocessed spectral data using a preset discrete wavelet transform strategy to obtain candidate effective feature signals. The feature extraction unit is used to extract feature signals containing significant discriminative features from the candidate valid feature signals based on a preset frequency feature combination strategy.
4. The serum Raman spectroscopy non-invasive pathological detection device according to claim 3, characterized in that, The preprocessing unit is specifically used for: The normalization subunit is used to normalize the qualified spectral data using a preset Z-score standardization to obtain normalized spectral data. The noise reduction and suppression subunit is used to perform noise removal and background suppression processing on the normalized spectral data to obtain the preprocessed spectral data.
5. The serum Raman spectroscopy non-invasive pathological detection device according to claim 3, characterized in that, The decomposition unit is specifically used for: A preset discrete wavelet transform strategy is used to perform multi-scale frequency decomposition on the preprocessed spectral data to obtain the first decomposition result.
6. The serum Raman spectroscopy non-invasive pathological detection device according to claim 1, characterized in that, Also includes: The training module includes: The sample acquisition unit reads spectral data samples from the preset database, preprocesses the spectral data samples, and obtains feature signal samples based on the preprocessed data samples. A partitioning unit is used to divide the feature signal samples into a training set, a validation set, and a test set based on a preset partitioning strategy. The training unit constructs a convolutional neural network and inputs the training set into the convolutional neural network to train and obtain initial model parameters; The validation unit is used to input the validation set into the convolutional neural network for performance evaluation based on the initial model parameters, and adjust the initial model parameters according to the performance evaluation results until the loss function of the validation set converges to obtain the optimal model parameters. The testing unit is used to input the test set into the convolutional neural network for model testing based on the optimal model parameter combination, and obtain the pre-trained convolutional neural network model when the test results meet the preset requirements.
7. The serum Raman spectroscopy non-invasive pathological detection device according to claim 6, characterized in that, The partitioning unit is specifically used for: Based on a preset random sampling strategy, a portion of samples are extracted from the feature signals to construct the test set, and the remaining samples are divided into the training set and the validation set using K-fold cross-validation.
8. The apparatus according to claim 1, characterized in that, Also includes: The display module is used to visualize the reasoning process and the pathological discrimination results through a preset graphical user interface, enabling clinicians to make non-invasive pathological diagnoses based on the displayed results.